Machine Learning Method for Prediction of Hearing Improvement After Stapedotomy

Otosclerosis is a localized disease of the bone derived from the otic capsule. Surgery is considered for patients with conductive hearing loss of at least 15 dB in frequencies 250 to 1000 Hz or higher. In some cases, the decision as to whether surgery (stapedotomy) should be performed is challenging. We developed a machine learning method that predicts a patient's postoperative hearing quality following stapedotomy, based on their preoperative hearing quality and other features. A separate set of regressors was trained to predict each postoperative hearing intensity on selected feature sets. For feature selection, the least absolute shrinkage and selection operator (Lasso) technique was used. Four models were constructed and evaluated: Lasso, Ridge, k-nearest neighbors, and random forest. The most successful predictions were made at air conduction frequencies between 1000 and 3000 Hz, with mean absolute errors of approximately 6 dB. Utilizing the nested CV method, the Lasso predictor achieved the highest overall prediction accuracy. This study presents the first stapedotomy result prediction method for operating surgeons using machine learning. The potential of audiogram estimation in predicting hearing recovery is demonstrated, offering an alternative to existing classification based models.